Pith. sign in

REVIEW 3 major objections 7 minor 26 references

taz2024full: Analysing German Newspapers for Gender Bias and Discrimination across Decades

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper presents a corpus of over 1.8 million German taz articles from 1980 to 2024 and uses it to show that men are consistently overrepresented in coverage, with a gradual shift toward balance after 2010.

desk verdict The corpus is the real contribution; the gender findings are a demonstration whose numbers are not yet trustworthy until the coreference step is validated. read the letter →

arxiv 2506.05388 v1 pith:QMPMT4CU submitted 2025-06-03 cs.CL

Stefanie Urchs , Veronika Thurner , Matthias A{ss}enmacher , Christian Heumann , Stephanie Thiemichen This is my paper
classification cs.CL
keywords Germannewspapercorpusgenderbiasmediarepresentationdiachronicanalysiscoreferenceresolutionsentimentactor-leveldiscoursetaz
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper presents taz2024full, a freely accessible corpus of over 1.8 million articles from the German daily taz published between 1980 and 2024, and argues it is the largest public German newspaper corpus currently available. As a demonstration, it applies an actor-level discourse-analysis pipeline to the whole corpus and reports that men are consistently overrepresented as named actors and as mentions across all four decades, although actor inclusion becomes nearly balanced after 2010. It also reports that sentiment toward women is slightly but consistently more negative than sentiment toward men throughout the entire period. The contribution is thus both a large linguistic resource for diachronic German-language research and a scalable method for measuring gender representation and framing over time.

What carries the argument

The carrying mechanism is the corpus itself combined with an adapted actor-level bias-detection pipeline. Named-entity recognition collects person names and generic person terms; coreference resolution links German pronouns to actors; a 70% pronoun-majority rule assigns each actor to one of two binary gender categories; sentences mentioning an actor supply sentiment scores from a German sentiment model; and pointwise mutual information, the log ratio of observed co-occurrence to expected co-occurrence, identifies adjectives most strongly associated with each gender. Yearly aggregation of these signals turns the raw corpus into a time series of representation and framing.

What would settle it

Take a stratified random sample of about 1,000 articles spanning 1980 to 2024, manually determine each named actor's gender from biography or context without using pronouns, and compare with the pipeline's pronoun-based labels; systematic disagreement in ambiguous German sentences, especially masculine defaults, would overturn the overrepresentation and sentiment-gap findings.

Watch

Extended reading notes

Core claim

The paper's central claim is that taz2024full is the largest publicly available corpus of German newspaper articles to date, containing over 1.8 million articles from the daily taz between 1980 and 2024. Applied to this corpus, an actor-level discourse-analysis pipeline reveals that men are consistently overrepresented in both actor counts and mention counts, though the share of women actors rises toward parity around 2010 while mention counts still lag. The paper also finds that sentiment toward women is slightly more negative than sentiment toward men in every year of the sample. These results are offered as a demonstration of the corpus's usefulness, not as a definitive study of German media bias.

Load-bearing premise

The results assume that German pronoun coreference and the 70% pronoun-majority rule correctly identify an actor's gender; if the coreference model systematically defaults ambiguous pronouns to masculine, the male overrepresentation and sentiment gap could be artifacts of the tool rather than properties of taz reporting.

Editorial extensions

If this is right

  • Researchers studying German-language change can use the corpus to trace vocabulary, topics, and framing over 44 years without negotiating license fees.
  • The demonstration establishes a reproducible baseline: any other German outlet, if released, can be run through the same pipeline and compared with taz.
  • The finding that a left-leaning paper still overrepresents men, especially in mention counts, sets a benchmark for what balanced media coverage would require.
  • Updating the corpus beyond 2024 can test whether the post-2010 trend toward balance continues or stalls.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One implication the authors leave implicit is that the remaining imbalance is concentrated in textual space: even when women appear as actors at near-parity, they receive fewer mentions, so future studies should weight exposure, not just presence.
  • Because taz is a single Berlin-based outlet, the corpus cannot speak for German media as a whole; a direct test would be to run the same publicly available pipeline on another German archive if licensing were solved.
  • The binary pronoun framing means non-binary actors are effectively excluded, so the measured two-way gap may understate representational skew in the full population of actors.
  • The sentiment result depends on a model trained on German newspaper language; comparing it with human ratings on a sample would clarify whether the small negative gap toward women is a property of the texts or of the model.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper introduces taz2024full, a corpus of 1,834,370 German newspaper articles from taz (1980–2024), claims it is the largest publicly available German newspaper corpus to date, and demonstrates its use by analyzing gender representation and sentiment over four decades. The analysis adapts an actor-level pipeline from Urchs et al. (2024), using spaCy NER, coreferee for pronoun coreference, a 70% pronoun-majority rule for gender assignment, and a German sentiment model. The authors report a consistent overrepresentation of men as actors and in mentions, a recent shift toward more balanced actor inclusion, and a persistently slightly more negative sentiment toward women. The corpus and code are released on Zenodo and GitHub, respectively.

Significance. If the corpus is as described, it is a valuable contribution to German-language NLP and computational social science: it offers a long, diachronic, single-source news corpus with open access, which is rare for German. The paper also publishes the full pipeline, which supports reproducible actor-level bias analysis. The authors are transparent about several limitations, including the single-source and left-leaning nature of taz, the difficulties of German coreference resolution, and the absence of a gold standard. However, the demonstration analysis that motivates the corpus's utility currently rests on unvalidated measurement choices; the manuscript's own Limitations section concedes that coreference resolution on German data remains problematic. This is not a reason to reject, but it requires substantive additional evidence before the empirical gender findings can be taken at face value.

major comments (3)
  1. [§4.2 and Figures 3–4] The central empirical claims—male overrepresentation and a persistent sentiment gap—depend entirely on the pronoun-based gender assignment described in §4.2, yet the method is not validated. The pipeline includes only actors for whom coreferee establishes pronoun coreference and labels an actor as woman or man only if more than 70% of associated pronouns are feminine or masculine. The manuscript does not report the share of actors left undefined, the distribution of pronoun counts per actor, any minimum pronoun count, or any comparison against a gold-standard annotation sample. The Limitations section itself states that coreference resolution on German data 'remains problematic.' If German coreference fails more often for feminine referents (e.g., because 'sie'/'ihr' are ambiguous), the observed male overrepresentation and the sentiment gap could be artifacts of differential measurement error rather than properties of taz reporting. Please add a validation study on a manually annotated sample stratified by gender, report the undefined-actor rate over time, and run a sensitivity analysis varying the 70% threshold and the minimum pronoun count.
  2. [§3.2 and §3.3] For a resource paper, the corpus construction is under-specified. The crawling procedure is described only as 'crawling publicly available content' between August and November 2024; there is no description of the starting URL list, crawling rules, handling of pagination or paywalled content, deduplication, or how the final article count of 1,834,370 relates to the full set of articles taz published over the period. The paper also notes that only publicly available content is included and that counts decline after 2007, but it does not quantify the gap between published and crawled articles. Without these details, the central claim that taz2024full is 'the largest publicly available German newspaper corpus' and spans 1980–2024 cannot be fully verified or independently reproduced. Please add a detailed data collection protocol, deduplication statistics, and, where possible, a comparison with the publisher's own archive counts.
  3. [§4.3 and Figure 4] The sentiment finding—that sentiment toward women is 'consistently slightly more negative' than toward men over the entire 44-year period—is presented without any uncertainty quantification. Figure 4 appears to show small differences, but the paper does not report effect sizes, confidence intervals, or statistical tests, so it is unclear whether these differences are distinguishable from noise across years and across the varying number of articles per year. Please add per-decade differences, standard errors or confidence bands, and at least a simple test of whether the female–male sentiment gap is significantly different from zero in each decade or overall.
minor comments (7)
  1. [§2] The reference to 'TüPP-D/Z corpus (?)' contains an unresolved placeholder; please either cite the original source or remove the entry. Also, the reference 'Sabine et al., 2004' should be 'Brants et al., 2004' to match the reference list.
  2. [§3.3 and Table 1] The units in Table 1 are ambiguous: 'token length' appears to mean characters per token, and 'sentence length' appears to mean tokens per sentence. Please clarify the units. Additionally, '6,944,197 unique tokens' should be 'unique token types' if that is the intended meaning.
  3. [§3.2] Please provide the exact crawling tool and configuration (e.g., Scrapy version, polite delay, retry policy) to support reproducibility.
  4. [§4.2] Specify the versions of spaCy, coreferee, and the German sentiment model ('oliverguhr/german-sentiment-bert') used in the pipeline, since coreference and sentiment performance can vary significantly across versions.
  5. [§4.2 and Equation (1)] For the PMI calculation, please clarify how P(x, y), P(x), and P(y) are estimated (e.g., counts over the whole corpus, smoothing), and whether the values are computed per actor or per gender category.
  6. [Figures 3 and 4] The figures are difficult to interpret as presented: please add axis labels, units, and legend entries, and consider adding confidence bands or error bars. The current black-and-white rendering makes the two series hard to distinguish.
  7. [Appendix A] The aggregated report listing has broken whitespace in several metric names (e.g., 't o t a l _ f e m i n i n e _ c o d e d _ w o r d s'); please fix the rendering.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the corpus claim and the gender measurements are empirical/self-contained, and the self-cited pipeline is an applied tool rather than a forced derivation.

full rationale

The paper's primary deliverables are a new corpus and a descriptive gender analysis. The corpus claim ('the largest publicly available corpus of German newspaper articles to date') is an external comparison, not a consequence of a definition. The gender findings in Section 4.3 are empirical counts and average sentiment values obtained by applying an NLP pipeline (NER, coreference resolution, pronoun-based gender labeling) to the corpus. Although the categories 'woman'/'man' are defined by thresholding associated pronouns (Section 4.2: 'If more than 70% of the pronouns associated with an actor are either feminine or masculine, the actor is categorised as woman or man, respectively'), the output proportions are not fixed by that definition: the pipeline could have returned balanced counts, and the observed male overrepresentation is a contingent measurement rather than a tautology. The paper's adoption of the same authors' prior pipeline (Urchs et al., 2024) is a methodological self-citation, but it is not load-bearing in the circularity sense: the present results are not asserted solely on the authority of that citation, and no 'prediction' in this paper reduces to a parameter fitted from the target data. The Limitations section explicitly concedes that 'co-reference resolution, particularly on German data, remains problematic due to the lower accuracy of current models for this language' and notes the absence of benchmarks. That is a validation gap and a correctness risk, not a self-definitional or fitted-input circularity. No equation in the paper is identical by construction to the reported gender imbalance, and no quantity is fitted to reproduce the observed overrepresentation or sentiment gap. Therefore no circular step can be exhibited.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central resource claim rests on the completeness and legal availability of the crawl; the demonstration analysis rests on pronoun-to-gender inference, coreference accuracy, and model validity. There is one explicit hand-chosen threshold (70%) and no invented theoretical entities. The largest-corpus claim would require a more systematic comparison to be fully verified.

free parameters (1)
  • pronoun majority threshold = 70%
    Actors are classified as woman or man when more than 70% of their co-referenced pronouns are feminine or masculine; the threshold is chosen without sensitivity analysis (Section 4.2).
assumptions (4)
  • domain assumption German coreference resolution links pronouns to the correct actor with sufficient accuracy.
    Section 4.2 uses coreferee for pronoun-to-actor linking but provides no accuracy evaluation on this corpus; systematic errors would change the gender counts.
  • domain assumption Pronoun usage is a valid binary proxy for gender.
    Section 1 restricts the analysis to binary genders because German lacks widely used non-binary pronouns; Section 4.2 assigns actors to woman/man when at least 70% of their pronouns match. This is an acknowledged modeling decision, not an empirical fact.
  • domain assumption SpaCy NER and the German sentiment model 'oliverguhr/german-sentiment-bert' work reliably on taz articles of 1980-2024.
    Section 4.2 relies on these models without a validation sample or domain adaptation check for historical newspaper text.
  • domain assumption The public taz archive crawl is representative of the full publication history.
    Section 3.3 notes article counts decline after 2007, possibly due to paywalled content, so the corpus may miss part of the published record; no completeness audit is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of taz2024full: Analysing German Newspapers for Gender Bias and Discrimination across Decades." pith.science (2026). https://pith.science/paper/QMPMT4CU

@misc{pith2026250605388,
  author       = {Pith},
  title        = {Pith review of: taz2024full: Analysing German Newspapers for Gender Bias and Discrimination across Decades},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QMPMT4CU}},
  note         = {Machine review of arXiv:2506.05388}
}
read the original abstract

Open-access corpora are essential for advancing natural language processing (NLP) and computational social science (CSS). However, large-scale resources for German remain limited, restricting research on linguistic trends and societal issues such as gender bias. We present taz2024full, the largest publicly available corpus of German newspaper articles to date, comprising over 1.8 million texts from taz, spanning 1980 to 2024. As a demonstration of the corpus's utility for bias and discrimination research, we analyse gender representation across four decades of reporting. We find a consistent overrepresentation of men, but also a gradual shift toward more balanced coverage in recent years. Using a scalable, structured analysis pipeline, we provide a foundation for studying actor mentions, sentiment, and linguistic framing in German journalistic texts. The corpus supports a wide range of applications, from diachronic language analysis to critical media studies, and is freely available to foster inclusive and reproducible research in German-language NLP.

Figures

Figures reproduced from arXiv: 2506.05388 by the authors.

Figure 1
Figure 1. Structure of the elements in the corpus, in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Number of newspaper articles and articles containing actors per year from 1980 to 2024. Years before [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the number of actors per article [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Sentiment towards the detected genders through the years. Looking at language use more closely, an analy￾sis of adjectives with the highest PMI association with women and men actors reveals that these de￾scriptors remain relatively stable over time and do not exhibit s…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 21 canonical work pages

  1. [1]

    https://www.dwds.de/ Der deutsche wortschatz von 1600 bis heute

    Berlin-Brandenburgischen Akademie der Wissenschaften . https://www.dwds.de/ Der deutsche wortschatz von 1600 bis heute. online

    work page

  2. [2]

    Su Lin Blodgett, Solon Barocas, Hal Daum \'e III, and Hanna Wallach. 2020. https://doi.org/10.18653/v1/2020.acl-main.485 Language (technology) is power: A critical survey of `` bias '' in NLP . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454--5476, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.acl-main.485 2020

  3. [3]

    o rklund, and Henrik Bj\

    Hannah Devinney, Jenny Bj\" o rklund, and Henrik Bj\" o rklund. 2022. https://doi.org/10.1145/3531146.3534627 Theories of “gender” in nlp bias research . In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT '22, page 2083–2102, New York, NY, USA. Association for Computing Machinery

    arXiv 2022

  4. [4]

    Danielle Gaucher, Justin Friesen, and Aaron C Kay. 2011. Evidence that gendered wording in job advertisements exists and sustains gender inequality. Journal of personality and social psychology, 101(1):109

    work page 2011

  5. [5]

    Anushka Gupta, Diksha Chugh, Anjum, and Rahul Katarya. 2022. Automated news summarization using transformers. In Sustainable Advanced Computing, pages 249--259, Singapore. Springer Singapore

    work page 2022

  6. [6]

    Daniel Jurafsky and James H. Martin. 2000. Speech and language processing. Prentice-Hall

    work page 2000

  7. [7]

    Marc Kupietz and Holger Keibel. 2009. The mannheim german reference corpus (dereko) as a basis for empirical linguistic research. Working papers in corpus-based linguistics and language education, 3:53--59

    work page 2009

  8. [8]

    Ken Lang. 1995. Newsweeder: Learning to filter netnews. In Proceedings of the Twelfth International Conference on Machine Learning, pages 331--339

    work page 1995

Show all 26 references
  1. [9]

    https://guides.loc.gov/chronicling-america/additional-features Newspaper datasets and api access

    Library of Congress . https://guides.loc.gov/chronicling-america/additional-features Newspaper datasets and api access . online

  2. [10]

    Asad Mahmood. 2017. https://www.kaggle.com/datasets/asad1m9a9h6mood/news-articles News articles . online

    2017

  3. [11]

    Camila M Mateo and David R Williams. 2020. More than words: a vision to address bias and reduce discrimination in the health professions learning environment. Academic medicine, 95(12S):S169--S177

    2020

  4. [12]

    Rishabh Misra and Jigyasa Grover. 2021. Sculpting Data for ML: The first act of Machine Learning

    2021

  5. [13]

    Andreas Nolda, Adrien Barbaresi, and Alexander Geyken. 2021. https://doi.org/doi:10.1515/9783110731514-018 Das ZDL-Regionalkorpus: Ein Korpus für die lexikografische Beschreibung der diatopischen Variation im Standarddeutschen , pages 317--322. De Gruyter, Berlin, Boston

    2021 doi

  6. [14]

    Thomas Proisl and Peter Uhrig. 2016. https://doi.org/10.18653/v1/W16-2607 SoMaJo : S tate-of-the-art tokenization for G erman web and social media texts . In Proceedings of the 10th W eb as C orpus Workshop ( WAC-X ) and the EmpiriST Shared Task , pages 57--62, Berlin. Associa...

    2016 doi

  7. [15]

    Martin Reisigl. 2017. https://doi.org/10.1007/978-3-658-10976-9_8 Sprachwissenschaftliche Diskriminierungsforschung , pages 81--100. Springer Fachmedien Wiesbaden, Wiesbaden

    2017 doi

  8. [16]

    Brants Sabine, Stefanie Dipper, Peter Eisenberg, Silvia Hansen, Esther König, Wolfgang Lezius, Christian Rohrer, George Smith, and Hans Uszkoreit. 2004. TIGER: Linguistic Interpretation of a German Corpus . Journal of Language and Computation, 2:597--620

    2004

  9. [17]

    Dietmar Schabus, Marcin Skowron, and Martin Trapp. 2017. https://doi.org/10.1145/3077136.3080711 One million posts: A data set of german online discussions . In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIG...

    2017

  10. [18]

    Anne Schiller, Simone Teufel, Christine Stöckert, and Christine Thielen. 1999. Guidelines für das Tagging deutscher Textcorpora mit STTS . Technical report, Institut für Maschinelle Sprachverarbeitung, Universität Stuttgart

    1999

  11. [19]

    Andrew Schwartz, and Dirk Hovy

    Deven Santosh Shah, H. Andrew Schwartz, and Dirk Hovy. 2020. https://doi.org/10.18653/v1/2020.acl-main.468 Predictive biases in natural language processing models: A conceptual framework and overview . In Proceedings of the 58th Annual Meeting of the Association for Computatio...

    2020 doi

  12. [20]

    Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2019. https://doi.org/10.18653/v1/P19-1159 Mitigating gender bias in natural language processing: Literature review . In Proceeding...

    2019 doi

  13. [21]

    taz. 2018. https://taz.de/taz-Geschaeftsfuehrer-als-Zeitzeuge/!vn6036515/ taz-geschäftsführer als zeitzeuge: Es begann in einem kalten ladenlokal | taz.de . online

    2018

  14. [22]

    taz. 2024. https://taz.de/Verlag/Die-Seitenwende/!v=165b2a3c-a94a-4114-8616-1b254db7ced2/ Die seitenwende | taz.de . online

    2024

  15. [23]

    https://wortschatz.uni-leipzig.de/de Willkommen beim wortschatz-portal

    Universität Leipzig , Sächsischen Akademie der Wissenschaften zu Leipzig , and Instituts für Angewandte Informatik . https://wortschatz.uni-leipzig.de/de Willkommen beim wortschatz-portal . online

  16. [24]

    Stefanie Urchs, Veronika Thurner, Matthias A enmacher, Christian Heumann, and Stephanie Thiemichen. 2024. https://doi.org/10.18653/v1/2024.gebnlp-1.8 Detecting gender discrimination on actor level using linguistic discourse analysis . In Proceedings of the 5th Workshop on Gend...

    2024 doi

  17. [25]

    Zimmerman

    Candace West and Don H. Zimmerman. 1987. https://doi.org/10.1177/0891243287001002002 Doing gender . Gender & Society, 1(2):125--151

    1987 doi

  18. [26]

    Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2019. https://proceedings.neurips.cc/paper_files/paper/2019/file/3e9f0fc9b2f89e043bc6233994dfcf76-Paper.pdf Defending against neural fake news . In Advances in Neural Inf...

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.